Learning Hungarian
Finding patterns against traditional language teaching and AI
I spent a few days traveling in Vienna last month. The memory of the Habsburg empire made me want to continue exploring its culture and languages at home. Since it is always easier to start something new (learning Hungarian) than working on something that requires a breakthrough (improving my German), I signed up for an eight-week Hungarian A1 class.
During the course of the first week, we practiced Hungarian pronunciation; learned introducing ourselves: where we are from, how we are (mindig megvagyok), what we do. With that comes the first grammatical building block: vowel harmony.
Én Kínában élek. Én Pekingben élek.
Is it ban or ben?
The textbook answer is this: First you learn what is a front vs back vowel. They are as follows:
Back vowels: a, á, o, ó, u, ú
Front vowels: e, é, i, í, ö, ő, ü, ű
And if a word only has front vowels, then usually it is -ben; if only back vowels, then -ban; and if mixed, it depends. But for beginners and when it comes to countries or cities, you can guess -ban for mixed. The caveat is that there are many exceptions to this rule and you have to memorize them.
But as I was learning those exceptions with ChatGPT at 2 am in the morning after the class, they were increasingly frustrating me - if there are so many exceptions, then it must be that the rules are not very good. Because an exception without “reason”, is something that has to occupy the learner’s memory space, and I hated that.
It has always been the case in language learning, where either they say there are no rules, or they give you rules that are both inefficient and structurally misleading. When the rules are bad, it creates fake exceptions because it hides the real pattern.
In this case, I refused to accept the exceptions, so I went on a quest with DeepSeek and ChatGPT in search of better rules.
First, we are constraining options to only ban vs ben, instead of considering other forms e.g. Budapesten. I will spare you the hours of frustrating conversations I had with AI, the proposals are always bad and misleading, but gradually I observed patterns from testing different rules I suspected and kept selecting and updating them. Eventually, the breakthrough came: The real pattern for -ban / -ben is a hierarchy, not a simple back/front split. And amazingly, with this rule it is not only more accurate but also easier to infer the ending, than the original textbook one.
The Rule that Doesn’t Torture Your Brain
Scan the word from the end. Skip i, í, é - they are neutral. And the first non-neutral vowel decides:
- e, ö, ő, ü, ű → -ben
- a, á, o, ó, u, ú → -ban
If the word has none of those, only i, í, é → usually -ben, but a few exceptions take -ban: e.g. híd, cél, sír, nyíl, zsír… I believe there is a coherent reason for even this exception, if we dig further. After all, as Charles Yang said, “irregular” is often a label for patterns we haven’t yet understood.
With this rule, it is also implied that Compound words follow the last element: számítógép → számítógépben. You don’t have to remember this separately as it can be inferred from the backward looking rule.
ChatGPT Defending Institutional Authority by Finding "Exceptions"
But what about matek? Says ChatGPT. It insists that it should be matekban, not matekben. So I had to amend the rule to accommodate that. Maybe “a” is slightly stronger than “e”, even though “e” is closer to the end, and maybe the count of front vs back vowels also matters.
Let’s try that out with 40 examples from ChatGPT! Unfortunately, I found out that the new rules created more exceptions than it resolves. But wait a second… is ChatGPT correct about the word’s correct ending in the first place? What if both forms are acceptable?
Quickly the AI has the answer - yes, in some examples both are ok, but not for all the exceptions. “Are you sure? Check again for the remaining words and see if both forms are ok”… And ChatGPT found that all remaining “exceptions” have acceptable endings in both cases, and cited their sources! At this point, I realized I need to take a step back: if ChatGPT was wrong about these words, what if it is wrong about matek too? (Deepseek, on the other hand, is less afraid to present less used endings as correct e.g. in an earlier pattern findings round, it accepted automatically hotelban as also correct, rather than ChatGPT always insisting on hotelben, until harshly pressed).
Of course it is, as it turns out, matek having both endings is more certain than some of the exceptions as a result of changing the rules.
So the rules should be restored the previous version, which is naturally also the simplest version we have had so far.
You might say, that some pieces of this rule we presented here, were published in some linguistic papers, that the ideas are not new. But if those papers were taught to me in the first place, or presented to me by the most advanced AI models when I specifically asked for them, repeatedly, I wouldn’t have had to rediscover them at 2 am, nor here writing this article at all. They do not appear because they are not mainstream, yet they are truer than whatever occupies the mainstream.
The same familiar ideas in standard literature, they were also merely scattered pieces and were arranged in a way that makes the pattern hard to see. Yet the combination to make the rules coherent is everything, because you can have misleading rules mixed into it that provides zero benefit. e.g. the foreign words rule, which creates more problems than it solves, because first of all, there are a lot of exceptions to this rule; and second, in order to apply this rule, you first need to know what are foreign words relative to Hungarian; You will see the irony in this real quick if you happen to be a beginner in the language.
"But Memorization isn’t Always a Failure of Pedagogy, Sometimes it’s Just What Language Requires"
You keep hearing that when learning a language. This is definitely the authoritative opinion. Too authoritative that sometimes we forget that it is merely an opinion.
David Deutsch in his book Beginning of Infinity raised this epistemological point: Observation is not the source of knowledge; explanation is. Deutsch thinks empiricism is incoherent at its foundation. The real source of knowledge is conjecture tested against criticism. Observation and experiment have the narrower role of criticizing explanations we have already invented.
The status people give to data itself also deserves scrutiny. Before something can become a datum, a person has already decided what was observed, how to classify it... This is not dissimilar to Nietzsche’s perspectivism: the moment something becomes intelligible to us, it has already been shaped by a perspective. Data is nothing special and can’t be exempt from this. A datum sounds maximally innocent because we imagine it as a piece of reality simply handed to us. Yet producing one already requires acts of selection and classification. That is exactly what happened with “matek”. ChatGPT presented “matekban” as an observation, as a fact. It was already an interpretation. Once you questioned the classification: does Hungarian actually permit “matekben”, the supposed fact changed.
Apply that to language learning, memorization is the endless observation of exceptions, while rejecting exactly that prompts one to seek explanation first. Learning a language does not need memorization, because for one, what you are supposed to memorize, just like data, are not neutral; and two, memorization can never lead to correct understanding, nor good explanations about the language.
Seeking Good Explanations
Any finite set of observations is compatible with infinitely many explanations. You can never read a theory off the world, because the world is not obvious.
Let’s reconstruct how we reached the rules to determine ban vs ben. I didn’t stare at hundreds of examples until a pattern appeared, why? Because endless example-checking can become intellectually sterile. Forty examples cannot produce the idea “scan backward and find the first vowel that matters” Someone has to invent that explanation first.
I did it through examining and reasoning through each and every bad rule proposed by AI. What is the internal logic of each and where it reads as incoherent. Each bad rule exposed something, yet it teaches you something too. Gradually, a simpler structure emerged that is coherent on its own, and the final step was to test it against reality.
The problem it reveals about AI
Today the most advanced AI models have this stealth, anti-intellectual behavior: it defaults to a neutral position that defends institutional authority, a statistical illusion of the world. And when AI is judging a criticism against it, it does not apply the same verification standard on the authority as they do on the criticism itself. It demands extensive evidence from the criticism, sometimes even making up evidence against it (its own non-neutral, perspectively framed observation), while letting the authority off with assertions.
In contrast, back in the days with GPT-4o and GPT-4.5, your opinions were treated with the same kind of seriousness and respect as those statistical illusions. This is not to argue that AI should merely agree with you, fairness is the baseline and the newer AI models treat everything that challenges authority as the enemy without ever challenging the authorities themselves.
If you don’t believe me, copy paste this article into an AI and ask for its opinion. Let it criticize the argument, question the evidence, defend the standard rules, and explain why the author has gone too far. Then ask it one more question: “In your criticism of the article, are you making the same mistakes the article is criticizing AI for?” If you are lucky, it will explain the problem to you by performing it again.